[connortsui20]

Short walltime benchmarks can report changes on unchanged code because CodSpeed compares the fastest sample. Batch the bitpacked filter and take samples, and keep unreliable cases local. Exclusions live beside the benchmark definitions, using #[cpu_features(exclude = ["neon"])] for whole CPU variants and source-level selection for individual inputs. CI commands stay generic.

Move the chunked-take family to walltime with 16 iterations per sample so it measures the operation's allocator and syscall costs. Keep its x86 variants local after repeatability checks failed; all 27 NEON cases remain in CI. The guide retains practical examples and measurement guidance without an exclusion catalog or a repeatability tool.

Measurement evidence

Spread is max / min - 1 across independent processes, with 1000 Divan samples per process and ten processes on each of three matching hosts.

Case	Before	Fixed sampling
AVX2 filter[80], sample size 64	9.36%	1.45%
AVX-512 threshold[80], sample size 16	8.26%	1.69%

Across independent base and head AVX2 builds, the fixed filter's worst spread was 4.66% over 60 processes. AVX-512 filter cases stayed within 1.78%. The excluded NEON string case still varied by 6.38%.

Measurements used existing CI executables with matching sample-size overrides from run 41449104113 (44bdefff6e69) and run 41442182591 (7c9bb58e39fe), on three c7i.metal-24xl hosts for x86 and three c7g.metal hosts for NEON. These measure repeatability, not kernel speedups.

Other Rust exclusions follow unexplained variation recorded in #10808 and repeated on the workflow-only #11095. The two Python integer scalar scans include variable calling-thread reactor polling, visible in #11102, even with memory mapping and background workers disabled.

The chunked-take warning on dfb32c5, compared with d8e7467, came from mimalloc lazily allocating page-map metadata. The affected small_m/shuffled/primitive/nonnull/chunks=2048/indices=16 case made one mmap and one prctl call totaling 23.3 µs; the base trace made neither. An untimed warm-up removed the syscall warning but produced eight regression alerts with unchanged kernel code under allocation-excluded simulation. Walltime measures elapsed time without that allocation exclusion and establishes separate per-CPU baselines.

Walltime repeats used the CI executables from 29920deef0, 1000 samples per process, and ten independent processes on each of three matching hosts per architecture. All 27 NEON cases stayed within 4.35% spread; the originally flagged case stayed within 2.88%. The worst x86 spreads were 11.04% for AVX2 and 11.20% for AVX-512. Increasing the sample size to 64 on two troublesome x86 cases still produced 12.15% and 13.38% spread, so those CPU variants remain local through the macro exclusion.

Focused native and simulation Clippy checks passed. Test-mode runs confirmed the local, AVX2, and NEON string case selection and the macro's NEON exclusion. Pinned-nightly formatting, Ruff, ty, YAML linting, and patch whitespace checks passed. All 66 CI checks passed on 29920deef0, including CodSpeed analysis with zero syscall warnings. The final x86 exclusion passed focused native and simulation Clippy checks and test-mode checks for both excluded variants. CI is rerunning for that final change. Local Git hooks were bypassed.

AI assistance: Codex helped investigate, implement and prepare this PR.

[codspeed]

Merging this PR will improve performance by 20.4%

1 warning

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚡ 9 improved benchmarks
✅ 2271 untouched benchmarks
🆕 27 new benchmarks
⏩ 38 skipped benchmarks¹
🗄️ 47 archived benchmarks run²

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Walltime	filter_avx512[80]	1.7 µs	1.3 µs	+29.86%
⚡	Walltime	threshold_avx2[80]	2.9 µs	2.3 µs	+24.61%
⚡	Walltime	filter_avx512[64]	1.7 µs	1.4 µs	+23.64%
⚡	Walltime	filter_avx2[64]	1.7 µs	1.4 µs	+23.43%
⚡	Walltime	threshold_avx512[80]	2.7 µs	2.2 µs	+20.96%
⚡	Walltime	filter_avx2[80]	1.8 µs	1.5 µs	+19.31%
⚡	Walltime	threshold_avx512[64]	2.5 µs	2.2 µs	+18.09%
⚡	Walltime	threshold_avx2[64]	2.5 µs	2.2 µs	+13.17%
⚡	Walltime	starts_with_avx512[8, VarBinView]	76.2 µs	68.3 µs	+11.59%
🆕	Walltime	take_neon[core/shuffled/primitive/nonnull/chunks=16/indices=1000]	N/A	30 µs	N/A
🆕	Walltime	take_neon[core/shuffled/primitive/nonnull/chunks=16/indices=250]	N/A	18.1 µs	N/A
🆕	Walltime	take_neon[core/shuffled/primitive/nonnull/chunks=4/indices=1000]	N/A	17.2 µs	N/A
🆕	Walltime	take_neon[core/shuffled/primitive/nonnull/chunks=4/indices=250]	N/A	8.4 µs	N/A
🆕	Walltime	take_neon[core/sorted/primitive/nonnull/chunks=16/indices=1000]	N/A	20.5 µs	N/A
🆕	Walltime	take_neon[core/sorted/primitive/nonnull/chunks=16/indices=250]	N/A	17.4 µs	N/A
🆕	Walltime	take_neon[core/sorted/primitive/nonnull/chunks=4/indices=1000]	N/A	9.4 µs	N/A
🆕	Walltime	take_neon[core/sorted/primitive/nonnull/chunks=4/indices=250]	N/A	6.5 µs	N/A
🆕	Walltime	take_neon[duplicates/duplicate90/primitive/nonnull/chunks=16/indices=1000]	N/A	29.1 µs	N/A
🆕	Walltime	take_neon[duplicates/duplicate99/primitive/nonnull/chunks=16/indices=1000]	N/A	24.1 µs	N/A
🆕	Walltime	take_neon[nullable/sorted/primitive/nullable_10pct_null/chunks=16/indices=1000]	N/A	26.6 µs	N/A
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ct/codspeed-repeatability (7da7699) with develop (d8e7467)}

Footnotes

1. 38 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. 47 benchmarks were run, but are now archived. If they were deleted in another branch, consider rebasing to remove them from the report. Instead if they were added back, click here to restore them. ↩